What is Gibbs’ Canonical Distribution?

Kevin Davey

Abstract

Although the canonical distribution is one the central tools of statistical me-
chanics, the reason for its effectiveness is poorly understood. This is due in part
to the fact that there is no clear consensus on what it means to use the canon-
ical distribution to describe a system in equilibrium with a heat bath. In this
paper, I examine some traditional views as to what sort of thing we should take
the canonical distribution to represent. I argue that a less explored alternative,
according to which the canonical distribution represents a time ensemble of sorts,
has a number of advantages that rival interpretations lack.

1 Introduction.
One striking thing about the machinery of modern statistical mechanics is how
effective it is. It is able to explain and predict a great number of facts about sys-
tems in equilibrium, and even some facts about systems out of equilibrium. More
striking, however, is the fact that the reason for the effectiveness of this machinery
is so poorly understood, despite so much effort by physicists, mathematicians and
philosophers on this very issue.
Let us focus for now on the use of the canonical distribution to calculate quan-
tities associated with a system in equilibrium with a heat bath. What justifies this
procedure? Answering this question requires carefully untangling several distinct
issues. First, we must ask what it means to describe the state of a system using
the canonical distribution. Armed with the answer to this question, we must ar-
gue that we are right to describe systems in equilibrium with heat baths using the
canonical distribution. Finally, we must argue that when we do correctly describe
a system using the canonical ensemble, we are justified in expecting the value
of a macroscopic observable on the system to be that given by integrating the
observable over phase space using the canonical measure.
Some small progress can be made towards justifying the use of the canonical
ensemble by carefully distinguishing these questions, and treating them one by
∗
Thanks to Brandon Fogel, Nick Huggett and John Norton for their feedback and suggestions.

1
one. In what follows, I provide a sketch of how to begin this task, focusing mainly
on the question of what it means to describe the state of a system using the
canonical distribution.

2 What is the Canonical Ensemble?

2.1 Equilibrium
We want to ask what it means to describe a system in equilibrium with a heat
bath using the canonical distribution. It will be useful to begin by asking what it
means to be in equilibrium with a heat bath at all. (For now, we take the concept
of a heat bath as unproblematic.)
The concept of equilibrium has a disparate set of meanings, and we need to
fix on one. Sometimes when we talk about a system being in equilibrium, we
mean that the system is in a particular state – generally one of the macrostates
described by the mathematical formalism of classical equilibrium thermodynamics
– at a given point in time. I call this a static notion of equilibrium, insofar as it
is generally used to determine whether a system is in equilibrium (possibly with
something else) at a given time. This is not the only way in which we can think
of equilibrium. In fact, when we talk about a gas being in equilibrium with a heat
bath, we generally do not have a static conception of equilibrium in mind. The gas
will exchange energy with the heat bath in such a way that it will occassionally find
itself out of static equilibrium. So the following rough definition will suffice for our
purposes: a system is in equilibrium with a heat bath if, as it evolves over time, it
spends most of its time in the macrostate (or macrostates) associated with static
equilibrium, only occasionally fluctuating into non-equilibrium macrostates. This
is a temporally extended notion of equilibrium, insofar as saying that a system
is in equilibrium with a heat bath in this sense is now making a claim about the
behavior of the system over time.
Insofar as heat baths are supposed to be infinite, it may well turn out that
any finite system placed in contact with a heat bath will behave in the way just
described, if studied over a sufficiently large time scale. Thus, it may well turn
out that any finite system in contact with a heat bath is in equilibrium with the
heat bath, in the sense just defined. This should cause no confusion, so long as
it is understood that there are different concepts of equilibrium, and a system in
one sort of equilibrium need not be in another type of equilibrium.

2.2 Traditional Interpretations.

Let us return to the canonical distribution. According to Gibbs, the state of a
system at equilibrium with a heat bath at temperature T is properly described by
the ‘canonical’ distribution:
e−βH(X)
ρ(X, t) = (1)
Z

2
 where X is a point in phase space, H is the Hamiltonian and Z the partition
function. (For simplicity, we suppose that the Hamiltonian is time-independent,
and that our system is constrained to lie in some finite volume of physical space.)
But what does it mean to say of a system that it may be described by the canonical
distribution (1)? I shall take it for granted that describing systems in contact
with heat baths using the canonical distribution involves making some sort of
probabilistic or statistical claim about such systems. But precisely what sort of
probabilistic or statistical claim could this be?
One approach is to think of the canonical distribution as giving us epistemic
probabilities. In its simplest form, the claim here is that that when we know a
system is in equilibrium with a heat bath, it is rational to assign the degree of
belief ρ(X)dX to the proposition that the system is in region dX at time t, where
ρ is the canonical distribution. But a claim like this is surely implausible. I might
have a strong belief that the number of particles in a gas lies in a certain range (say,
between 1.1×1023 to 1.2×1023 particles), but for each integer N in this range I will
presumably assign a very low probability to N being the exact number of particles
in the system. The canonical distribution, however, will assign probability 1 to
the claim that the gas has exactly N particles, where N is the actual number of
particles described in the Hamiltonian. In this case, the canonical distribution
gives a probability of 1 where my epistemic probability is almost 0. A more
sophisticated proponent of the epistemic interpretation of statistical mechanics
might have something to say about this sort of worry,1 though whether related
worries arise is a question I am happy to leave open. For the purpose of this
paper, I will focus instead on the idea that describing a system with the canonical
distribution involves making an objective, non-epistemic probabilistic or statistical
claim about the system at hand.
More specifically, according to the approaches that I do want to consider,
describing a system with the canonical distribution involves making two obser-
vations: first, that the system is a member of a particular ensemble (or ‘set’ of
systems), and second, that it is appropriate (in some sense to be specified) to de-
scribe this ensemble using the formula (1). To pursue this approach, we must first
construct an ensemble of systems, and then explain in what sense it is appropriate
to describe this ensemble with the formula (1).
There are a few ways to carve up possible ways of proceeding. One thing we
might do is try to construct our ensemble out of actually occurring systems with
Hamiltonian H, in contact with a heat bath at a given temperature T . We call
this an actualist approach:
Actualist Approach: The ensemble should consist of some appropri-
1
One might try to argue that our epistemic probabilities are given by weighted sums of canonical
distributions, each involving a slightly different Hamiltonian, and that because the observables that
matter have (roughly) the same expected value when calculated with this weighted sum as when cal-
culated with the canonical distribution involving the actual Hamiltonian, we may take the canonical
distribution to represent our epistemic state well enough for the purpose of calculating the expected
macrostate of the system.

3
 ately chosen set of actual (i.e., physically instantiated) systems with
the Hamiltonian H, in contact with a heat bath at temperature T .
Given an ensemble constructed in accordance with this requirement, we might then
claim that the statistics of the ensemble be (approximately) given by the canonical
distribution. More precisely, we might claim that the proportion of systems in the
ensemble occupying a fixed small region of phase space is (approximately) equal
to the value of (1) in that region of phase space.2 We call this an (approximate)
identity requirement:
(Approximate) Identity Requirement: The proportion of systems in the
ensemble in a fixed small region of phase space is (approximately) equal
to the value of the canonical distribution in that region of phase space.
But the combination of an actualist approach with an approximate identity re-
quirement is not promising at all. Presumably, there are systems that have only
found themselves in contact with heat baths a small number of times in the his-
tory of the universe. Only a tiny region of their phase space will ever have been
instantiated, and perhaps only a tiny region of their phase space will ever be in-
stantiated. The canonical distribution will typically not describe the distribution
of any set of instantiations of such systems.
In response to this, let us stick with the actualist approach, but try to relax
the approximate identity requirement. To motivate our replacement for the ap-
proximate identity requirement, consider a coin that has been thrown 100 times.
Take the hypothesis:
Hypothesis: If the coin were to be thrown continually, it would come
up heads about half the time, and tails about half the time.
Suppose that the coin has come up heads 2 times and tails 98 times. In such a
case, we would be inclined to say that Hypothesis has been strongly disconfirmed
– where by ‘strong disconfirmation’ I mean (loosely speaking) a type of discon-
firmation that is especially severe. If the coin came up heads 35 times and tails
65 times, we might say only that the hypothesis has been disconfirmed, but not
strongly so. If the coin came up heads 48 times and tails 52 times, we would be
inclined to say that Hypothesis has not been disconfirmed at all. Exactly how
statistical confirmation or disconfirmation works, and whether and to what extent
it needs to draw on a notion of prior probability, will not be important to us. The
important point is the scientific commonplace that claims about the distribution
of an ensemble can, in certain cases at least, be strongly disconfirmed by claims
about a sample.
Let us suppose now that we have some procedure for sampling systems with
Hamiltonian H, in contact with a heat bath at temperature T . The result is a
(finite) set of systems distributed through phase space in some particular way.
Consider now the hypothesis:
2
We assume that the regions of phase space in question are sufficiently small that the value of (1)
does not change significantly in them.

4
 Hypothesis: If we continually sample systems with Hamiltonian H in
contact with a heat bath at temperature T , we will get an ensemble
with the property that the proportion of systems in the ensemble in a
small region of phase space is (approximately) equal to the value of (1)
in that region of phase space.
This hypothesis is capable of being disconfirmed – and even strongly disconfirmed
– by an actual (finite) sample of systems with Hamiltonian H in contact with a
heat bath at temperature T . So instead of the approximate identity requirement,
we impose the following requirement:
Non-Disconfirmation Requirement: Hypothesis is not strongly discon-
firmed by our actual (finite) sample of systems with Hamiltonian H in
contact with a heat bath at temperature T .
The Non-Disconfirmation requirement can handle the case of uninstantiated or
rarely instantiated systems much better, insofar as we typically do not take hy-
potheses about distributions to be strongly disconfirmable by very small (or even
empty) samples.
The Non-Disconfirmation requirement is very modest – perhaps too modest
to really be useful. It is interesting, then, that even a requirement this modest
gets us into trouble. To see how, let us consider the details of the process by
which we sample systems. Suppose we sample the microstate of a system with
Hamiltonian H in contact with a heat bath at temperature T only at the moment
the system is first placed in contact with the heat bath. The problem is that many
systems are way out of static equilibrium when first placed in contact with a heat
bath. Because of this, the method of sampling just described may consistently
produce a disproportionate number of systems in microstates that are extremely
rare according to the canonical distribution (1). As such, the hypothesis that
continual sampling will result in an ensemble whose statistics are given by the
canonical distribution will be strongly disconfirmed.
Perhaps it is unfair to sample the microstates of systems only at the moment
they are placed in contact with heat baths. But if we just include all of the
microstates of the system while it is in contact with a heat bath, the out-of-static-
equilibrium microstates may still be disproportionately represented. (Imagine for
instance a system with a very long relaxation time, such that up until now, such
systems in contact with heat baths at a given temperature have spent almost all
of their lives in regions of phase space that are extremely rare according to the
canonical distribution.) It also will not do to sample the microstate the moment
the system reaches static equilibrium, as the canonical distribution predicts energy
fluctuations that will not be present if we only sample microstates with fixed en-
ergy. The absence of such energy fluctuations will count as strong disconfirmation
of Hypothesis, particularly in cases in which the occurrence of such fluctuations
is decently probable.
The most promising strategy is to sample the microstate of the system in
contact with the heat bath only after some prespecified amount of time – some sort

5
 of ‘relaxation time’ – appropriate to the system at hand. But I am skeptical as to
whether this sort of approach can work. Highly contingent facts about the world,
including facts about the way in which we choose to prepare systems, have a way
of showing themselves in any such ensemble. For instance, consider a system that,
at a given temperature, has multiple macroscopically distinguishable equilibrium
states. Any system with potentially variable anisotropic structure, such as a
crystal, will do. For a more artificial example, consider a 1-dimensional Ising ring
with no external magnetic field, i.e., a system of N particles with Hamiltonian:
N
X
H = −J σi σi+1 ,
i=1

in which each σi takes on the value ±1, and σN +1 = σ1 . At low temperatures, this
system organizes itself into large, macroscopic blocks of 1s or −1s. Because the
Hamiltonian is unchanged under the transformation σ → −σ, for each equilibrium
macrostate at a given temperature there is a distinct equilibrium macrostate at the
same temperature in which each macroscopic chunk is replaced by one of opposite
sign. At sufficiently low temperatures, these equilibrium macrostates are clearly
distinguishable.To make the argument that follows as clean as possible, let us
suppose that systems with the given Hamiltonian do not occur naturally, but are
always manufactured by us. It is perfectly possible for us, having manufactured
such a system, to place it in contact with a heat bath at a certain cold temperature
T much lower than that of the environment in which it is produced, if and only
if the systems is in a particular equilbrium macrostate – for instance, iff the
largest macroscopic block consists of 1s, rather than −1s. When placed in contact
with the heat bath, such a system is very likely to ‘relax’ into a particular sort
of macrostate – in this case, a macrostate in which it remains the case that the
largest macroscopic block consists of 1s rather than −1s. As a result, insofar as we
only sample systems after this relaxation time, our sample will consist primarily
of microstates in a very specific region of phase space – namely, the region of
phase space corresponding to this particular sort of macrostate. This ensemble
will look extraordinarily unlikely, given the canonical probability distribution. If
we place 1000 such systems in contact with heat baths, and essentially all of them
end up with the largest macroscopic block consisting of 1s rather than −1s, it will
be as if a coin thrown 1000 times came up heads essentially each time, which is a
situation with extraordinarily low probability. In this case, we would have to say
that Hypothesis is strongly disconfirmed.3
3
This is not to deny that if any such system remained in contact with the heat bath long enough,
there would eventually be large energy fluctuations, after which the system might find itself into one
of the other possible macrostates at the given temperature. But this will typically occur only after
an astronomically long time. So if we try to get around the problem by insisting that we only sample
systems after sufficiently large energy fluctuations have occured, all our ensembles will typically be
empty, and our description of these ensembles with the canonical distribution starts to border on
vacuity. Whether obstacles like this can be satistfactorily surmounted is something I am happy to leave

6
 I think that difficulties like these show that the entire actualist approach is
in error. Insofar as our interpretation of the canonical distribution involves iden-
tifying an ensemble, I suspect that the ensemble will have to be populated not
just with actual physical systems, but also with counterfactual systems. This is
in agreement with Gibbs, who denied that his ensembles had any sort of ‘objective
existence’, saying instead that they were ‘creations of the imagination’. (See p.188
of [8].) But what counterfactual ensemble should we pick?
The traditional view, championed by Gibbs, was to simply posit an imaginary
collection of systems distributed in accordance with the canonical distribution.
According to this approach, when one describes a system using the canonical
distribution, one is making a claim about an imaginary ensemble of which the
particular system is a member. There are many well known problems with this
approach. For instance, the actual system is a member of all sorts of imaginary
ensembles – so why the privilege the canonical ensemble when it comes to making
predictions?4 I think the fact that problems like this are so difficult suggests that
this way of thinking of the canonical distribution does not quite get things right,
and that it makes sense to look for a more ‘naturally occuring’ counterfactual
ensemble with which to work.

2.3 The Time-Ensemble

One particular proposal seems particularly natural. The rough idea is this – take
a system in contact with a heat bath, and imagine allowing it to stay in contact
with the heat bath indefinitely. Look at the distribution ψ of the set of microstates
through which the system then passes. To describe a system in equilibrium with
a heat bath using the canonical distribution is just to say that ψ is the canonical
distribution.
Although I think this idea is basically right, it must be formulated very care-
fully in order to avoid being clearly wrong. The most obvious way to define a
distribution φ on the set of microstates through which our system passes is as
follows: focus at first on the microstates through which the system passes in a
fixed interval of time from t = 0 to t = T . Let dX be a very small region of space
space – for simplicity, we assume it is a ball centered around a point X of phase
space – and let τ (dX) be the amount of time that the system spends in dX in the
interval [0, T ] . Then we would like to have:
τ (dX)
ψ(X)|dX| ≈
T
With this in mind, we define:
τ (dX)
ψ(X) = lim .
|dX|→0 T |dX|

open – at any rate, these obstacles are challenging enough that it is worth considering alternative points
of view.
4
For extremely thorough discussions of this problem, though mainly with respect to the microcanon-
ical rather than canonical ensemble, see [4] and Chapter 5 of [10].

7
 where we imagine the radius of the open ball dX going to 0. The problem is that
it may be shown to follow from this that ψ(X) = 0 almost everywhere.5 This is
not a useful distribution with which to work.
This sort of situation is not unfamiliar. In his Lectures on Gas Theory [1],
Boltzmann considers the case of a homogeneous gas in a vessel. He defines a
function f (~v , t) – the so called ‘density of velocities’ – by requiring f (~v , t)|d~v | to
be the number of particles in a (fixed) region of space whose velocities lie in the
region d~v at time t. Let us suppose we are dealing with point particles, and let us
suppose that the region d~v is a sphere centered around a fixed velocity ~v . Suppose
N is the number of particles in the region of space in question with velocity ~v at
time t. Then we must have:

lim f (~v , t)|d~v | = N.

|d~v |→0

If N 6= 0, no possible value of f (~v , t) can make this equation true, and if N = 0,

every value of f (~v , t) makes this equation true. So our definition has broken down.
Boltzmann’s solution to think of d~v as a fixed, tiny cell of velocity space cen-
tered around ~v . Because d~v is fixed, it is not the sort of thing we can send to 0.
Consequently, the equation
f (~v , t)|d~v | = N
defines a unique value of f (~v , t). Boltzmann’s idea is that we should choose our
cells d~v in such a way that the resulting distribution of velocities f (~v , t) is more
or less continuous – so that when we move from a cell of velocity space to an
‘adjacent’ cell, f (~v , t) does not change too much. That it is possible to choose
cells d~v in this way is a substantive assumption Boltzmann is happy to make. So
long as this assumption is made, the quantity f (~v , t) is well-defined enough for
Boltzmann to work with.
We mimic this move. We divide phase space into small cells. Let dX be the
cell containing X. We then define

τ (dX)
ψ(X) = . (2)
T |dX|

The idea here is to choose the cells in such a way that ψ is more or less continuous
– so that when we move from one of these cells to an adjacent one, ψ(X) does not
change too much. That it is possible to do this is a substantive assumption that
5
In a finite amount of time, the path our system traces through phase space will generally consist of
finitely many continuous curves, each of finite length. It may be shown to follow from this that the set
of points through which the system does not pass is an open set, possibly plus some finite set of points,
and that the set of points through which the system does pass is of measure 0. If we take a point X
through which the system does not pass, it follows that, with finitely many exceptions, there will be
some open ball centered around that point through which the system does not pass. And so when dX
is sufficiently small, we will have τ (dX) = 0. Thus, for sufficiently small dX, τ (dX)/(T |dX|) = 0, and
so it follows from our definition that ψ(X) = 0.

8
 we simply take for granted in what follows. So long as this assumption is made,
the quantity ψ(X) is well-defined enough to work with.
With all this in mind, we formulate the following:
Main Definition I (Tentative): For a system to be in a state described
by the canonical distribution is for it to be the case that, for sufficiently
large T , the distribution ψ defined by (2) is approximately equal6 to
the canonical distribution.
(We leave open the possibility that different coarse grainings may be required at
different times.) It is only reasonable to talk about approximate equality in our
definition for two reasons – first, the function ψ is coarse grained, and so cannot be
identical with the canonical distribution, and second, there is no guarantee that
in any finite time the system will have explored all the regions of phase space in
exactly the correct proportions.7
A natural generalization of the previous definition is possible:
Main Definition II (Tenative): For a system to be in a state described
by the distribution ψ is for it to be the case that, for sufficiently large
T , the distribution ψ defined by (2) is approximately equal to ψ.
There are several reasons why I am only prepared to call such definitions ‘ten-
tative’. I shall outline one. There are a number of ways we can think of a heat
bath. According to one point of view, a heat bath is in a determinate microstate
at any point in time, and so at the moment our system is placed in contact with
a heat bath there are a variety of microstates in which the heat bath could be.
Consequently, there will be different distributions ψ depending on the particular
initial state of the heat bath. So rather than making a claim about a single ψ,
our definitions should make a claim about the most probable distribution ψ (in
some appropriately defined sense.) Whether this modification is necessary is not
clear to me. I will not try to decide this issue here, but will simply call the def-
initions ‘tentative’, to acknowledge that their details may need to be altered to
accommodate different ways of thinking about a heat bath.
Let us put all this to the side for now, and make some general remarks. In these
definitions, our ensemble is the set microstates through which the system would
pass were it left in contact with the reservoir indefinitely. To say that a system
can be described by the canonical distribution is then to make a statistical claim
about the (probable) structure of this time-ensemble. Note that one advantage of
working with these definitions is that the focus from the start is on time averages.
If time-averages are so important – as the huge amount of effort directed towards
understanding the relationship between time and ensemble averages suggests –
6
For our purposes, it suffices toR say that two functions f1 and f2 on phase space are approximately
equal iff the phase space integral dX|f1 (X) − f2 (X)| is very small.
7
One might wonder whether the distributions defined in (2) converge (in some sense) to the canonical
distribution as T → ∞. A requirement like this is possible, but because it does not add anything to
the present discussion, I omit it for now.

9
 it seems much more efficient to just define the canonical ensemble as a time-
ensemble of sorts. This is not to suggest that by redefining words we can avoid
any of the major philosophical problems with equilibrium statistical mechanics.
All the old problems still exist, though many of them will need to be formulated
in different language. The hope, however, is that better definitions can give us a
better perspective on what these problems are, and how deep they run.
To see an example of an old problem in a new form, note that to describe a
system with the canonical distribution (as defined in Main Definition I) is now to
make a substantive assertion about the system. It is not a-priori obvious that it is
ever correct to describe any system with the canonical distribution, let alone the
multitude of systems to which the canonical distribution is actually applied. This
is a very serious issue to which we shall return shortly.
One final remark: according to our definition, to describe a system with the
canonical distribution is to make a claim about the (very) long term behavior
of the system. As such, the values of observables calculated using the canoni-
cal distribution are to be thought of as time-averaged expectation values. The
time-averaged expectation value of an observable can be quite different from the
value of the observable at any moment of time. However, supposing we can show
that the variance of the observable over time is small, we can at least assert
that the observed value will probably be approximately equal to the time-averaged
expectation value.8 The observed value probably being approximately equal to
the time-averaged expectation value is compatible with the observed value be-
ing radically different from the time-averaged expectation value. But so long as
we realize that the canonical distribution only gives us the probable value of an
observable9 , there is no real problem with the fact that the actual value of the ob-
servable might be very different, unless we insist on demanding from the canonical
ensemble something that it is simply not designed to give.

3 Heat Baths and Canonical States.

In the previous section, we gave a tentative definition for what it means for a
system to be in a state described by the canonical distribution. Our next question
is whether systems in equilibrium with heat baths really do have states described
by the canonical distribution. Does the canonical distribution accurately describe
the set of microstates through which a system passes when it is in contact with a
heat bath?
This is a tremendously difficult question. Some of these difficulties are of a
mathematical sort, but others are more conceptual. For instance, we are interested
8
An analagous move is made in [9], albeit with respect to ensemble averages and the microcanonical
distribution.
9
Here, we are talking about probability derived from the time-ensemble in the way described. There
may be other notions of probability at play, due to the fact that the system in question will be a part of
other ensembles or collectives. The canonical ensemble makes no claim to tell us anything about those
other ensembles.

10
 in studying a system interacting with a heat bath – but precisely what sort of
system is a heat bath? A heat bath must contain an infinite amount of energy
if it is to be possible that it produce arbitrarily large energy fluctuations in the
system with which it interacts, as the canonical distribution predicts. In addition,
a heat bath must be able to absorb or give up an arbitrarily large amount of heat
without changing its own temperature. A heat bath is therefore an idealization.
The question, then, is not what the microscopic structure of actual heat baths
are, for there are no such things. Instead, the question is what sort of model we
should use for heat baths.
The basic idea with which most authors begin is that a heat bath should be a
system containing a countable infinity of particles spread out over a well defined,
infinite region of space (where these conditions are, of course, not intended to
be sufficient.) For instance, one might take a heat bath to be an infinitely long
chain of harmonic oscillators. Provided the initial conditions of these particles are
chosen correctly, such a system will contain an infinite amount of energy. Because
heat capacity is an extensive quantity, the heat capacity of such a system will also
be infinite. Thus, we have what is at least a prima facie candidate for a heat bath.
Let us be even more specific. Take an infinitely long chain of harmonic oscilla-
tors, ordered like the integers. Take some finite subset of adjacent oscillators (e.g.,
the 1st, 2nd, .., nth.) This finite sub-chain of oscillators will be our system, and
the remainder of the oscillators will be our heat bath. Does this system actually
behave in the way that we would expect a system in a heat bath to behave? This
mathematical question has been studied in great detail. One of the pioneering
works is Ford, Kac, and Mazur [5]; subsequent important contributions include
Heurta and Robertson [6] [7], Davies [3], and Tegmark and Yeh [11].
Unfortunately, the upshot of this body of literature is difficult to assess. First,
on the terms on which much of this literature proceeds, the theorems proven there
show that our candidate for a heat bath does not behave in the desired way,
at least in general. But second, it is not clear that the theorems proved in these
papers are the sorts of theorems one would actually want to prove if one’s goal was
to show that the heat-bath-candidate behaves in the desired way. I will address
each of these points in turn. My conclusion is that the question of how to best
model a heat bath is unresolved. Because of this, the question of whether a system
in equilibrium with a heat bath should be described by the canonical distribution
(invoking Main Definition I) remains wide open.
First, let us consider the results of this literature on its own terms. In this
literature, it is assumed that we can talk about a probability distribution that
describes the system at any given time, and that this distribution undergoes the
usual time evolution determined by the Hamiltonian. This way of thinking about
what it is to describe a system with a probability distribution goes against our
Main Definition II, according to which the probability distribution describes a
time ensemble rather than an ensemble at an instant10 , but it is in this sense that
10
The fact that we may want to describe the heat bath with a probability distribution at a given
instant does not contradict the idea that the canonical distribution first and foremost makes a claim

11
 we consider the literature on its own terms. In [6], Huerta and Robertson study
a system with the Hamiltonian
n=+∞
X p2n k(xn+1 − xn )2
H= [ + ]
n=−∞
2m 2

They assume that the initial positions and momenta of particles both in the system
and in the bath are given by uncorrelated Guassians. In particular, they write:
Y {exp −[xn (0) − un ]2 /2α2 } exp {−[pn (0) − vn ]2 /2δ 2 }
ρ({x}, {p}, t = 0) = ·
α(2π)1/2 δ(2π)1/2
n∈{1...N }
Y exp {−xn (0)2 /22 } exp {−pn (0)2 /2ζ 2 }
× ·
(2π)1/2 ζ(2π)1/2
n∈{1...N
/ }

where it is assumed that α, δ, , ζ are all non-zero. Note that the above formula
allows the initial conditions of the particles in the system to be skewed. They
then show that
lim hpr (t)pr+1 (t)i =
6 0
t→∞
(where h·i denotes the phase space average with respect to the measure ρ at time
t.) Because there are no correlations between the momenta of different particles
in the canonical distribution, it follows that the system does not relax into the
canonical distribution.11 Thus, we cannot think of an infinite chain of oscillators
with initial conditions given by uncorrelated Gaussians as an effective model of a
heat bath.12
The argument of Tegmark and Yeh [11] is a little more general, but broadly
similar conclusions are reached. In the conclusion of their paper, when they dis-
cuss distributions proportional to e−H/kT (such as the canonical distribution),
they claim that ‘no completely solvable heat bath model has ever been found that
explicitly evolves multiparticle (n > 1) systems into such states’ (p. 359).
There is (almost) no questioning the mathematics behind the calculations of
Huerta and Robertson or Tegmark and Yeh. What is less clear is what these
calculations actually show. Both pairs of authors assume that what needs to
be determined is whether the probability distribution associated with the initial
state of a particular open system evolves into the canonical distribution as t → ∞.
Indeed, this is the approach of a large body of literature to which these authors’
papers belong.
But it is not clear what is demonstrated by the fact that an initial probability
distribution does (or does not) evolve into a particular final probability distribu-
tion as t → ∞. For according to Main Definition II, to describe a system using
about time ensembles.
11
The argument in §6 of [6] applies only to particles in the bath (a restriction imposed ‘for simplicity’),
but it is easy to generalize this argument to cover particles in the system.
12
Huerta and Robertson do show that some features of equilibrium are obtained in the long run,
such as equipartition of energy.

12
 a probability distribution ρ is to make a claim about how the system evolves as
t → ∞. Given these definitions, it does not make sense to interpret the relaxation
of a system into equilibrium as the evolution of one probability distribution to
another.13 Indeed, it does not make sense to talk about a system in contact with
a heat bath changing its state ρ at all. Insofar as our goal is to intepret and justify
the machinery of equilibrium statistical mechanics, this is not worrisome.
The net result of all this is that is difficult to interpret the calculations of Huerta
et. al as telling us anything meaningful about the set of microstates through which
a system passes when it remains in contact with a putative heat-bath. Because
of this, it is difficult to draw any conclusions from this literature about what
does or does not count as a good model of a heat bath. Any serious attack of this
question needs to examine the way in which particular initial systems (rather than
ensembles of initial systems) move through phase space when placed in contact
with a putative heat bath (or ensemble of putative heat baths.) To my knowledge
such matters remain unexplored, and we do not explore them here. The important
point is that the manner in which one approaches the physics question of how to
model a heat bath is determined in part by the manner in which one approaches
the philosophical question of what kind of thing the canonical distribution is.

4 Concluding Remarks.
The question of why we are right to describe systems in equilibrium with heat
baths using the canonical distribution remains open. Even once this problem is
addressed, other difficult issues then require attention. For instance – why it is
possible to treat a physical system in equilibrium with a large, finite reservoir
of heat as if it were in contact with an infinite reservoir of heat? And why it
is permissible to use the canonical distribution to describe systems that are not
in contact with any sort of reservoir of heat, as is often done? (The fact that
the microcanonical distribution and canonical distribution approach each other
in the thermodynamic limit is often claimed to justify the use of the canonical
distribution for closed systems, but it is far from clear that this argument is
convincing.) Regardless of all this, one moral that I hope can be drawn from this
paper is that thinking about what sort of thing the canonical distribution is can
help to clarify the mathematical questions that must be addressed in order to then
justify its use. This is at least a modest sort of progress.
13
Note, however, that it does make sense to talk about a change in ρ when a heat bath is replaced
with a different bath, insofar as one can talk about the states through which the system would have
passed were it to have remained in contact with the original heat bath, as well as the states through
which it will pass now that it is in contact with a new heat bath. Thus, there is the possibility of
talking about the way in which the state changes during certain quasi-static processes.

13
References
[1] Boltzmann, L., Lectures on Gas Theory, Dover Publications, 1964.
[2] Callendar, C., ‘Reducing Thermodynamics to Statistical Mechanics: The
Case of Entropy’, The Journal of Philosophy, Vol. 96, No. 7 (July 1999),
348-373.
[3] Davies, E. B., ‘The Harmonic Oscillator in a Heat Bath’, Commun. Math.
Phys., Vol 33 (1973), 171-186.
[4] Earman, J., and Redei, M., ‘Why Ergodic Theory Does Not Explain the
Success of Statistical Mechanics’, British Journal for the Philosophy of
Science, Vol. 47, (1996), 63-78.
[5] Ford, G., Kac, M., and Mazur, P., ‘Statistical Mechanics of Assemblies of
Coupled Oscillators’, Journal of Mathematical Physics, Vol 6, No. 4 (April
1965), 504-515.
[6] Huerta, M and Robertson, H., ‘Entropy, Information Theory and the Ap-
proach to Equilibrium of Coupled Harmonic Oscillator Systems’, Journal
of Statistical Physics, Vol 1, No. 3 (1969), 393-414.
[7] Huerta, M and Robertson, H., ‘Approach to Equilibrium of Coupled Har-
monic Oscillator Systems II’, Journal of Statistical Physics, Vol 3, No. 2
(1971), 171-189.
[8] Gibbs, J. Willard., Elementary Principles in Statistical Mechanics, Ox Bow
Press, 1981.
[9] Malament, D.,and Zabell, S., ‘Why Gibbs Phase Averages Work – The
Role of Ergodic Theory’, Philosophy of Science, Vol 47, 339-349.
[10] Sklar, L., Physics and Chance, Cambridge University Press, 1993.
[11] Tegmark, M. and Yeh, L., ‘Steady States of Harmonic Oscillator Chains
and Shortcomings of Harmonic Heat Baths’, Physica A, 202 (1994), 342-
262.

14